On 12.07.2012, at 00:43, Benjamin Herrenschmidt wrote:
On Thu, 2012-07-12 at 00:35 +0200, Alexander Graf wrote:
quoted
quoted
On Thu, 2012-07-12 at 00:25 +0200, Alexander Graf wrote:
quoted
Or just free up another register early on, like here.
=20
If you're going to do that, you want to measure the impact on null
syscall performance though.
=20
That should hold true for any change in that code, no?
=20
Yes, but adding a gpr save is more invasive (you have to also load it
back later on & put it on the stack).
What's the usual cache line size like on these boxes? If you align it =
properly with r10 and r11 on the paca so that all 3 registers are in the =
same cache line, the load should be almost for free, no?
Alex